Local AI builder weighs DGX Spark, A100 rig

Hot take: this is less about brand and more about whether they want convenience or maximum usable GPU for the money.

• –The DGX Spark option wins on simplicity, integration, and lower build risk, but it is still constrained by the bandwidth and architecture tradeoffs of an all-in-one box.
• –The A100 80GB route is the more serious local compute play if the adapter setup is stable, because raw VRAM and mature CUDA support matter a lot for training and larger inference workloads.
• –If the goal is mostly solo hobby work and experimentation, the strongest decision factor is whether the user can tolerate hardware/compatibility hassle in exchange for more flexible compute.
• –A third path would be used datacenter GPUs or a multi-GPU setup, but that only makes sense if power, cooling, chassis, and interconnect constraints are already solved.